This update is all about reliability: fewer dead ends while linking Telegram, fewer streaming failures in planning chat, and more resilient agent naming generation.
Telegram Linking That Survives Real-World Sign-In
Linking flows can get messy (redirects, multiple browsers, expired tokens). We hardened the end-to-end experience so you spend less time retrying and more time actually using your agents.
- Cross-browser safe linking - Linking uses a stable
stateparameter so the process works even if your sign-in flow changes the URL - Token preservation during sign-in - Your linking session is maintained through authentication so you do not have to start over
- Friendly failure page - If something goes wrong, you get a dedicated error page with a clear reason and the exact next step (usually: send
/loginagain)
Cleaner Telegram Chat Output
When agents are streaming tool-driven output, Telegram should still feel readable.
- Improved tool streaming - Tool-related output is less likely to get garbled or partially rendered
- More reliable linking backend - Fewer edge-case failures during linking and message delivery
Planning Chat: Migrated To A Dedicated Service
Planning chat moved to a more robust service layer, which reduces the number of "it works here but not there" bugs and makes streaming/tool execution more predictable.
- More predictable streaming - Responses are assembled correctly even when tools run across multiple turns
- Stricter message parsing - Tool result message parts are handled more defensively, reducing mysterious crashes
- Legacy code path removed - Less branching logic means fewer regressions and faster fixes
Agent Naming: Fewer Timeouts, Better Retries
Agent creation should not feel like waiting on a fragile stream.
- Reduced stream timeouts - Naming suggestions start and complete more consistently
- Improved SSE error handling - Better behavior when partial streams interrupt or reconnect
- Richer diagnostics - When failures do happen, we can debug them faster (which translates to faster fixes for you)
Default Skills Load Correctly In Production
- Vercel path fix - Default skill resolution was adjusted so agents reliably load their built-in guidance in production environments